Hyperscalers' Custom AI Chips
Rebellions
Custom AI chips turn compute from a purchased input into a controlled part of the cloud product. AWS splits training and inference across Trainium and Inferentia, Google has spent years building TPU pods into its ML stack, and Microsoft is now using Maia for both model serving and internal AI workloads. That lets hyperscalers tune cost, power, networking, and software together instead of paying NVIDIA’s full stack margin on every deployment.
-
The real advantage is not just cheaper chips. AWS wraps Trainium and Inferentia in Neuron, EC2, and SageMaker, Google exposes TPUs through Cloud TPU pods, and Microsoft is adding an SDK around Maia. In practice, customers buy a usable training or inference system, not a bare processor.
-
Each hyperscaler aims at its own highest volume jobs. Inferentia is for serving models, Trainium is for training, Google now has an inference first TPU generation in Ironwood, and Microsoft says Maia powers OpenAI models, Bing, GitHub Copilot, and ChatGPT workloads. The chip roadmap follows internal demand first.
-
This changes the opening for independent chip startups. NVIDIA still has the broadest software moat through CUDA and TensorRT, but hyperscalers increasingly reserve their biggest fleets for in house silicon. That pushes companies like Rebellions toward sovereign AI projects, telecom, finance, and regional clouds that want alternatives without building chips themselves.
The next step is wider verticalization. Hyperscalers will keep pairing custom silicon with their own compilers, networking, and managed AI services, while Chinese clouds deepen domestic stacks around chips like Hanguang and Ascend. That leaves the market splitting into a few full stack cloud platforms and a separate lane for specialized vendors serving customers outside those ecosystems.